Cisco's Agent Tournament: 1200 Engineers, and the Winner Wrote 1220 Tokens Instead of 1430
Cisco turned AI agent training into a competition: in 2026, 1200 engineers played in the AWS AI League tournament. In qualifying, the four best agents all cleared the entire map, and the one that wrote less won: 1220 output tokens versus 1430 for second place, a gap of 16 points out of almost 16,000. Think of two support bots with equally correct answers: the chatty one doesn't win. When everyone solves the task, count the tokens the agent writes itself.
AI-processed from AWS Machine Learning Blog; edited by Hamidun News
Any company rolling out AI agents knows the problem: everyone has access to the technology, but only a few people can confidently build products on it. Slides and webinar recordings barely close that gap. Cisco tried a different approach: it turned training into a tournament with a live leaderboard.
In 2026, 1200 engineers took part. The event also produced a useful look at what a winning agent looks like. Suppose you have two support bots that answer equally correctly.
One replies with a short sentence, while the other writes three paragraphs and recaps its own steps. At high request volumes, the difference in tokens turns into real money. That effect decided the tournament.
If you build agents, the article is worth reading for its practical takeaways: count the cost of every message, don't multiply agents unnecessarily, and don't make the model do what ordinary code can do. The material was published on the AWS blog, so the promotional angle is obvious, but the technical conclusions hold up without any tie to the cloud.
How the AWS AI League tournament works
AWS AI League has three steps. First comes a two-hour workshop with AWS experts. Then participants build a solution on their own and submit it to a leaderboard, where an automated grader responds after every attempt. Finally, there is a game-show-style finale on stage. Cisco ran two tracks: building agents on Amazon Bedrock AgentCore (the AWS platform for building and running agents on any model and framework) and fine-tuning models in Amazon SageMaker Studio.
In the agent track, a program navigates a grid maze to a treasure chest. Along the way it meets tasks that resemble real work: filtering inappropriate content, executing code, finding information on the web, and extracting structured data. Amazon Bedrock Guardrails handled safety, and AgentCore memory handled the agent's memory.
Here is the scale: at Cisco Transformation Fest 4.0, qualifying ran in four streams. The top five from each advanced to the semifinal, and of the 20 semifinalists, three reached the final, which had three live rounds. The winner, Cisco engineer Nishant Aluhonnu, placed first at every stage. In the final he scored 27,008 points in total and won a pass to AWS re:Invent 2026. A year earlier, Cisco engineer Hemant V. won the global final at re:Invent 2025, and Transformation Fest itself has reached about 2000 engineers since 2022.
What the winning agent showed
You pay for what the model writes. The score included a token bonus: 1000 minus the number of output tokens divided by the number of tasks visited. Output tokens are everything the agent generated itself.
In qualifying, the four best agents solved the whole map, and the one that wrote less won: 1220 output tokens versus 1430 for second place. The gap was 16 points out of almost 16,000. Nishant built nine variants of one agent, changing one parameter at a time.
He found that tool responses and the system prompt did not affect the score, while everything the agent wrote, including tool-call arguments, did. Each call added roughly 52 tokens of fixed overhead. The same logic applies in real life: generation is sequential and usually costs several times more than reading, and in an agent loop the costs accumulate at every step.
A long, structured prompt, which can also be cached, is cheap. A chatty agent is expensive.
One agent instead of five. The typical setup for this task is a main supervisor agent on AgentCore that delegates navigation, calculation, and web search to narrow subagents. For many products, that is a sound approach. The winner got by with a single agent and no subagents. His prompt opened with a short table: task type, and the right way to solve it in one turn. Every handoff between agents is a message the model has to write, and therefore pay for in tokens. The takeaway is not that subagents are bad. Choose an architecture based on the cost of coordination, and when that cost is high, a flat design can win.
The final and practical takeaways
In the final, participants could not change the code, tools, or prompts; only a short hint before each round was allowed. So everything came down to how well the agent made decisions on maps it had not seen before. Nishant lost the first round but won the second by a wide margin: he ignored a field strewn with cheap coins and went after a single expensive task. The system is not infallible, but its value logic worked where a habit of collecting everything in sight would have lost.
What to take away in practice. Before adding a new agent to your design, calculate how many tokens each handoff between agents costs. Move anything deterministic into ordinary code, and keep a short routing table in the prompt. The caveats matter too: a tournament metric with a token bonus is artificial, and in production an error on a hard task usually costs more than an extra 200 tokens. This is also a report from an AWS partner, with no independent measurements. Still, if you need to build up a team's skills quickly, the "build, break, fix, and compare on a leaderboard" format is clearly more effective than lectures. Cisco has made it annual.
- Nov 20, 2026Amazon SageMaker Now Flags Idle GPUs and Rising Latency on LLM Endpoints on Its Own
- Nov 20, 2026AWS teaches AI agents to buy APIs in Marketplace and pay through a regular AWS bill
- Nov 20, 2026BMW Moves AI Agent Rules for 12,500 Developers Into CI/CD and a Shared Registry
- Nov 19, 2026Amazon and Google will hold 28.4% of Anthropic after its IPO, but without voting rights
Key Questions Answered
How the AWS AI League tournament works?
AWS AI League has three steps. First comes a two-hour workshop with AWS experts. Then participants build a solution on their own and submit it to a leaderboard, where an automated grader responds after every attempt. Finally, there is a game-show-style finale on stage. Cisco ran two tracks: building agents on Amazon Bedrock AgentCore (the AWS platform for building and running agents on any model and framework) and fine-tuning models in Amazon SageMaker Studio.
What the winning agent showed?
You pay for what the model writes. The score included a token bonus: 1000 minus the number of output tokens divided by the number of tasks visited. Output tokens are everything the agent generated itself. In qualifying, the four best agents solved the whole map, and the one that wrote less won: 1220 output tokens versus 1430 for second place. The gap was 16 points out of almost 16,000. Nishant built nine variants of one agent, changing one parameter at a time.
What are the key details on The final and practical takeaways?
In the final, participants could not change the code, tools, or prompts; only a short hint before each round was allowed. So everything came down to how well the agent made decisions on maps it had not seen before. Nishant lost the first round but won the second by a wide margin: he ignored a field strewn with cheap coins and went after a single expensive task. The system is not infallible, but its value logic worked where a habit of collecting everything in sight would have lost. What to take away in practice.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.